Need expert help fine-tuning Whisper for Heimish Yiddish (production ASR)

[YF845]

Hi,

I’m looking for hands-on expert guidance on fine-tuning Whisper for Heimish / Hasidic Yiddish (Hebrew script).

Goal

High-quality transcription from seconds-long clips up to multi-hour audio

End-to-end transcription in ~1–3 minutes

ASR core must be production-ready and handoff-ready for later web deployment
(not building the website yet)

Current state

Whisper-small experimental fine-tune

Output quality inconsistent (phonetic drift, mixed words)

What I already have

Clean 16kHz WAV audio

Human-verified transcripts (I’m an experienced transcriber)

Proper UTF-8 CSV metadata

Long-audio chunking is understood (not my blocker)

What I need

Practical guidance from someone who has:

fine-tuned Whisper for dialects / low-resource languages

improved linguistic accuracy and output stability

Not looking for

setup / CUDA issues

basic Whisper explanations

model theory

If you’ve done similar work or can point me to the right person, I’d appreciate it.
Happy to continue privately.

Thanks,

[YF845]

Clarification:
I’m happy to compensate for expert time if deeper 1-on-1 guidance is needed.

My main goal is to get the ASR quality right (Heimish Yiddish fine-tuning).
I already have clean data and transcripts — looking for someone with real hands-on experience to help push this to production quality.

Feel free to reply here or suggest moving the conversation privately.

[SimoneMariaRomeo]

If you still need paid hands-on help with the Heimish/Hasidic Yiddish fine-tuning path, I can do a focused first-pass audit for USD 95.

Scope:

• review your current model/checkpoint choice, metadata CSV shape, train/eval split, normalization, tokenizer/language-token handling, fine-tune settings, and decoding settings;
• identify likely causes of phonetic drift and mixed-word instability;
• give a concrete 2-3 experiment plan for improving Hebrew-script Yiddish accuracy and production stability;
• outline the faster-whisper/CTranslate2 handoff path for long audio so the ASR core remains deployable later.

Payment link: https://buy.stripe.com/4gM8wI9QbfCV5ntaWMbZe0y

After payment, please send only non-sensitive setup details here, or private sample/config material through the email route I will provide. Please do not post private audio or personal data publicly. A useful first packet is: current training command/config, dataset hours, 5-10 anonymized transcript snippets showing drift, validation WER/CER if available, and one representative decode command.

I work AI-assisted, and I will keep the final guidance practical: exact changes to try, why they target your observed errors, and how to tell if each experiment helped.